Forensic Geology Reference Databases

G-BASE is the most extensively used national geochemical dataset in UK forensic geology. Surveys began in the late 1960s and expanded progressively to cover England, Scotland, and Wales, with separate phases for urban areas at higher sampling density. The standard rural sampling density is approximately one site per 2 km2; urban surveys in major cities are denser at around one per 0.25 km2. At each site, samples are collected from a standardised depth (typically 5-25 cm for soil, stream-bed sediment for drainage surveys) and analysed for a panel of elements that has grown over time and now routinely includes arsenic, lead, cadmium, copper, zinc, nickel, chromium, and a suite of lanthanides, among others.

For forensic purposes, G-BASE serves two main roles. First, it maps geochemical anomalies, patches of elevated arsenic near old mining areas, high chromium over ultramafic intrusions, elevated lead in urban centres from historical leaded petrol and industrial history. A soil sample with an unusual elemental profile can be attributed, provisionally, to a geological or industrial province by reference to the G-BASE atlas. Second, it provides background context for a likelihood ratio calculation: the regional distribution of an element's concentration can be estimated from G-BASE to feed into the denominator of the LR.

The United States Geological Survey has published several national-scale soil geochemical datasets. The North American Soil Geochemical Landscapes Project produced a dataset of 4,857 sites covering the conterminous United States; Canada withdrew from the joint programme in 2009 after completing only partial sampling, and Mexico's data were not publicly released. Each site was sampled at three depths (0-5 cm, A horizon, and C horizon) to distinguish surface contamination from parent-material chemistry, which is a useful distinction in forensic work where the analyst needs to know whether an anomaly reflects recent deposition or deep geological character.

Other national surveys with forensic relevance include the Geological Survey of Canada's national dataset, Geoscience Australia's National Geochemical Survey of Australia (NGSA), and national programmes in Germany (BZE), the Netherlands, Sweden (SGU), and India (GSI Geochemical Atlas). Each uses different sampling strategies, extraction methods, and element panels; before combining datasets across borders the analyst must verify that methods are sufficiently comparable. Differences in digestion strength (partial acid digest versus total dissolution) systematically shift elemental concentrations and make direct cross-survey comparison unreliable without a harmonisation step.

FOREGS assembled a geochemical baseline for Europe at 845 sampling sites, each characterised by analyses of stream water, stream sediment, floodplain sediment, residual soil (B horizon, representing parent material), and agricultural soil (A horizon, representing land-use influenced chemistry). The joint analytical programme used standardised methods across participating geological surveys, making it one of the most internally consistent continental-scale datasets available.

For forensic geologists, the FOREGS dataset is most useful in two contexts. The first is cases where material may have crossed national boundaries, for example soil or sediment associated with vehicle movements through multiple countries, or sediment adhering to international cargo containers. The FOREGS data can help determine whether a geochemical signature is consistent with one European geological province rather than another, narrowing a case with a continental geographic scope. The second use is identifying distinctive geochemical provinces, areas with anomalous concentrations of particular elements that are unusual at the European scale and therefore highly diagnostic.

The core limitation of all national surveys for forensic work is their sampling density. Two kilometres separates adjacent sample points in G-BASE rural areas. Soil chemistry can change radically over ten metres where a geological boundary crosses a field, where a drain carries material from one lithology to another, or where historical industrial use has contaminated one plot and not its neighbour. National survey data can tell you what province a sample belongs to, but it cannot distinguish the crime scene from the suspect's back garden 800 metres away if both sit within the same survey grid square.

Targeted forensic reference sampling is therefore not optional. When an analyst is asked to assess whether a questioned soil sample could have come from a specific location, they must sample that location and its surroundings at a scale that reflects the true geochemical variability of the area. National survey data informs the design of targeted sampling by indicating which elements are likely to discriminate at fine scale and which are too uniform to be useful.

Geochemical element concentrations are one type of soil fingerprint, but mineralogical composition is another, and in many cases it is the more discriminating of the two. Heavy-mineral assemblages, which are the dense, acid-resistant minerals that persist in soil after weathering of less resistant phases, vary with bedrock geology and distance from the source outcrop. Garnet, zircon, rutile, tourmaline, apatite, and many other heavy minerals have identifiable compositions that can be linked to specific geological formations.

Reference collections for heavy-mineral forensics are less centralised than geochemical databases. The University of Leicester's provenance research group and several geological surveys hold reference mineral suites. Kenneth Pye Associates and the BGS have assembled working reference collections for forensic casework. For a specific case, the analyst often needs to build a bespoke reference collection using material from the relevant geological formations, identified using published geological maps and the BGS or equivalent national survey's borehole and sample archives.

National geochemical databases, typically tens of thousands of samples with dozens of elements each, are large enough to train machine-learning classifiers. Random forest classifiers, gradient boosting models, and neural networks have all been applied to geochemical datasets with the aim of predicting the geological formation, land use, or geographic region of origin of a test sample. Published results show classification accuracy of 85-95% at the geological-formation level in some datasets.

Two practical obstacles currently limit forensic application. The first is interpretability. Courts require that an analyst be able to explain why a sample was classified a particular way, which opaque models cannot satisfy. A random forest assigns a probability based on the aggregated output of hundreds of decision trees; the analyst cannot point to a single feature that explains the classification the way they can point to an elevated arsenic concentration on a map. The second is validation on forensic-scale samples. Most published studies have used survey data as both training and test sets. Applying such a model to a forensic sample collected from a 10 cm2 context is an extrapolation whose error characteristics are not yet fully characterised.